Anthropic Launches OSS Scanner, a Free AI Vulnerability Scanner

Anthropic has launched OSS Scanner, a free security tool that checks important open-source software projects for vulnerabilities, or security weaknesses. Projects choose whether to take part. The scanner uses Anthropic’s strongest AI models, including Claude Mythos, to examine the code in projects that have signed up. It sends reports directly to project maintainers without a person checking them first. This can help teams find problems faster, but maintainers must spend time checking the reports themselves. Eleven days after the October 8 launch, Anthropic reported 77 findings rated critical and 803 rated high by its models.

TL;DR

• Scale: As of October 19, 114 projects had been accepted from about 650 applications.
• Runs on: Anthropic provides the computing power for free. Maintainers provide a Dockerfile, a file that tells the system how to build the project. The security checks run in an isolated virtual machine without Internet access.
• Performance: Of 97 critical- and high-severity findings that were checked, 85 (88%) met Anthropic’s standard for reporting vulnerabilities.
• Best: Only 1 of the 97 checked findings was a false positive, meaning it was reported as a problem when it was not valid. wolfSSL says 72 of its 74 reports were valid.
• Worst: No person checks reports before maintainers receive them. Of the 97 checked findings, 11 repeated problems that were already known. Maintainers also report that some findings are given severity ratings that are too high.

What is Anthropic OSS Scanner?

OSS Scanner is a free Anthropic tool that regularly checks open-source code repositories for security problems using the company’s strongest AI models. Anthropic’s Frontier Red Team and Safeguards Research team run it together. Its design takes ideas from Google’s OSS-Fuzz, but uses AI language models instead of fuzzing tools that test software with many different inputs.

The main goal is to help with a growing review backlog. Over 6 months, Anthropic’s models found over 29,000 possible vulnerabilities. Human reviewers checked about 6,000 of them. Nearly 5,000 more reports were sent directly to maintainers who had asked to receive every finding, whether it had been confirmed or not. OSS Scanner turns that informal arrangement into a program that projects can choose to join.

How does OSS Scanner work?

Core maintainers can sign up by opening a pull request (PR) in anthropics/oss-scanner. A pull request asks for a change to a software project. It must add a file called projects/<project>/project.yaml, based on the project template.

• Required fields: repo, primary_contact, and the path to a Dockerfile.
• Optional fields: threat_model, pgp for encrypting reports, auto_ccs, homepage, and disabled.

The Dockerfile builds the project while network access is available. The virtual machine then switches to a network without Internet access for the security check. AI agents in the process check possible bugs again, investigate their root causes, and suggest fixes.

Each report includes an example that can reproduce the problem on its own, along with an explanation of the bug. When possible, it identifies the code change that introduced the problem by narrowing the search to a specific commit. A possible code fix is attached when one is available.

Maintainers can optionally provide a threat_model.md file. It describes the inputs an attacker might use, which parts of the code are outside the scope of the check, and how serious different problems should be considered. Without this file, the scanner estimates these details itself.

How accurate are OSS Scanner reports?

Anthropic asked its penetration testers to check 97 critical- and high-severity findings across 48 projects.

• 85 (88%) met the requirements for Anthropic’s coordinated vulnerability disclosure (CVD) process, which helps security teams report vulnerabilities responsibly.
• 11 described real problems but repeated issues that were already known or other findings from the scans.
• 1 was invalid.

Reports from project maintainers show a similar pattern. wolfSSL received 74 reports; all but 2 were valid, and 5 became CVEs. curl fixed CVE-2026-92392, a high-severity HTTP/2 upload bug called a use-after-free. This type of bug happens when software tries to use memory after it has been released. The vulnerability affects curl versions 8.12.0 to 8.22.0. It was fixed in version 8.23.0, released on October 14, 2026.

Anthropic connects these results to improvements in testing AI models. On the CyberGym benchmark, large language models (LLMs) improved from finding under 20% of vulnerabilities to finding over 85%. Anthropic’s broader CVD dashboard lists 6,157 disclosed vulnerabilities across 591 projects as of October 2. Of those, 516 have been patched.

OSS Scanner vs other open-source security programs

Sources: OSS-Fuzz disclosure guidelines, OpenAI Codex Security, DeepMind CodeMender.

Key Takeaways

• OSS Scanner gives important open-source projects free security checks using Claude Mythos.
• Reports reach maintainers without a person checking them first. Of 97 critical- and high-severity findings that were checked, 88% were valid.
• In the first 11 days, the scanner produced 77 findings rated critical and 803 rated high by its models.
• Projects join by submitting a PR with a Dockerfile. Security checks run without Internet access.
• There is no 90-day disclosure deadline for findings that have not been validated.

Read the launch announcement, the OSS Scanner FAQ, and the GitHub repository for more information.